-
Notifications
You must be signed in to change notification settings - Fork 6
/
Copy pathtrain_online.py
181 lines (133 loc) · 6.87 KB
/
train_online.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
import time
import cv2
from recoder import VideoRecorder
from logger import Logger
from replay_buffer import ReplayBuffer
from utils.train import set_seed_everywhere
from utils.environment import get_agent_types
from overcooked_ai_py.env import OverCookedEnv
from model.utils.model import *
from utils.agent import find_index
import hydra
from omegaconf import DictConfig
class Workspace(object):
def __init__(self, cfg):
self.work_dir = os.getcwd()
print(f'Workspace: {self.work_dir}')
self.cfg = cfg
self.logger = Logger(self.work_dir,
save_tb=cfg.log_save_tb,
log_frequency=cfg.log_frequency,
agent=cfg.agent.name)
set_seed_everywhere(cfg.seed)
self.device = torch.device(cfg.device)
self.discrete_action = cfg.discrete_action_space
self.save_replay_buffer = cfg.save_replay_buffer
# self.env = NormalizedEnv(make_env(cfg.env, discrete_action=self.discrete_action))
self.env = OverCookedEnv(scenario=self.cfg.env, episode_length=self.cfg.episode_length)
self.env_agent_types = get_agent_types(self.env)
self.agent_indexes = find_index(self.env_agent_types, 'ally')
self.adversary_indexes = find_index(self.env_agent_types, 'adversary')
# OU Noise settings
self.num_seed_steps = cfg.num_seed_steps
self.ou_exploration_steps = cfg.ou_exploration_steps
self.ou_init_scale = cfg.ou_init_scale
self.ou_final_scale = cfg.ou_final_scale
if self.discrete_action:
cfg.agent.params.obs_dim = self.env.observation_space.n
cfg.agent.params.action_dim = self.env.action_space.n
cfg.agent.params.action_range = list(range(cfg.agent.params.action_dim))
else:
# Don't use!
cfg.agent.params.obs_dim = self.env.observation_space[0].shape[0]
cfg.agent.params.action_dim = self.env.action_space[0].shape[0]
cfg.agent.params.action_range = [-1, 1]
cfg.agent.params.agent_index = self.agent_indexes
cfg.agent.params.critic.input_dim = cfg.agent.params.obs_dim + cfg.agent.params.action_dim
self.agent = hydra.utils.instantiate(cfg.agent)
self.common_reward = cfg.common_reward
obs_shape = [len(self.env_agent_types), cfg.agent.params.obs_dim]
action_shape = [len(self.env_agent_types), cfg.agent.params.action_dim if not self.discrete_action else 1]
reward_shape = [len(self.env_agent_types), 1]
dones_shape = [len(self.env_agent_types), 1]
self.replay_buffer = ReplayBuffer(obs_shape=obs_shape,
action_shape=action_shape,
reward_shape=reward_shape,
dones_shape=dones_shape,
capacity=int(cfg.replay_buffer_capacity),
device=self.device)
self.video_recorder = VideoRecorder(self.work_dir if cfg.save_video else None)
self.step = 0
def evaluate(self):
average_episode_reward = 0
self.video_recorder.init(enabled=True)
for episode in range(self.cfg.num_eval_episodes):
obs = self.env.reset()
episode_step = 0
done = False
episode_reward = 0
while not done:
action = self.agent.act(obs, sample=False)
obs, rewards, done, info = self.env.step(action)
rewards = np.array(info['shaped_r_by_agent']).reshape(-1, 1)
self.video_recorder.record(self.env)
episode_reward += sum(rewards)[0]
episode_step += 1
average_episode_reward += episode_reward
self.video_recorder.save(f'{self.step}.mp4')
average_episode_reward /= self.cfg.num_eval_episodes
self.logger.log('eval/episode_reward', average_episode_reward, self.step)
self.logger.dump(self.step)
def run(self):
episode, episode_reward, done = 0, 0, True
start_time = time.time()
while self.step < self.cfg.num_train_steps + 1:
if done or self.step % self.cfg.eval_frequency == 0:
if self.step > 0:
self.logger.log('train/duration', time.time() - start_time, self.step)
start_time = time.time()
self.logger.dump(self.step, save=(self.step > self.cfg.num_seed_steps))
if self.step > 0 and self.step % self.cfg.eval_frequency == 0:
self.logger.log('eval/episode', episode, self.step)
self.evaluate()
start_time = time.time()
self.logger.log('train/episode_reward', episode_reward, self.step)
obs = self.env.reset()
self.ou_percentage = max(0, self.ou_exploration_steps - (self.step - self.num_seed_steps)) / self.ou_exploration_steps
self.agent.scale_noise(self.ou_final_scale + (self.ou_init_scale - self.ou_final_scale) * self.ou_percentage)
self.agent.reset_noise()
episode_reward = 0
episode_step = 0
episode += 1
self.logger.log('train/episode', episode, self.step)
if self.step < self.cfg.num_seed_steps:
action = np.array([self.env.action_space.sample() for _ in self.env_agent_types])
if self.discrete_action: action = action.reshape(-1, 1)
else:
agent_observation = obs[self.agent_indexes]
agent_actions = self.agent.act(agent_observation, sample=True)
action = agent_actions
if self.step >= self.cfg.num_seed_steps and self.step >= self.agent.batch_size:
self.agent.update(self.replay_buffer, self.logger, self.step)
next_obs, rewards, done, info = self.env.step(action)
rewards = np.array(info['shaped_r_by_agent']).reshape(-1, 1)
if episode_step + 1 == self.env.episode_length:
done = True
if self.cfg.render:
cv2.imshow('Overcooked', self.env.render())
cv2.waitKey(1)
episode_reward += sum(rewards)[0]
if self.discrete_action: action = action.reshape(-1, 1)
dones = np.array([done for _ in self.env.agents]).reshape(-1, 1)
self.replay_buffer.add(obs, action, rewards, next_obs, dones)
obs = next_obs
episode_step += 1
self.step += 1
if self.step % 5e4 == 0 and self.save_replay_buffer:
self.replay_buffer.save(self.work_dir, self.step - 1)
@hydra.main(config_path='config', config_name='train')
def main(cfg: DictConfig) -> None:
workspace = Workspace(cfg)
workspace.run()
if __name__ == '__main__':
main()